| In recent years,with the rapid development of deep learning and computer vision technologies,the research of Synthetic Aperture Radar(SAR)automatic target recognition(ATR)has made great progresses.But it needs to be solved the practical application in the complicated working environment.This paper adopts deep learning techniques,combines transfer learning and attention mechanism theory to do the study of the SAR ATR theoretical work in depth,so as to improve the final effect.The following is a description of the specific research content of the SAR ATR algorithm in this article:Firstly,in terms of data processing,this paper proposes a constrained Naive GAN(CN-GAN)SAR image generation and enhancement algorithm based on the generative Adversarial Networks(GAN)technology so as to solve the problems of sparse SAR image sample size and low signal-to-Clutter-Noise Ratio(SCNR).CN-GAN effectively combines least-squares GAN(LSGAN)with image-to-image translation(Pix2Pix)methods,while improving the loss function of the generator to achieve the purpose of generating samples and enhancing samples.Experimental results show that CN-GAN achieves better image generation effect,effectively inhibits coherent spot noise of SAR images,enhances image quality,and improves the accuracy of target recognition.Secondly,in terms of network design,a SAR ATR algorithm combining deep transfer learning with attention mechanism is proposed in order to solve the long training problem and the poor model prediction problem when the noise interference is severe.The algorithm uses deep convolutional network to achieve end-to-end target recognition,and adopts the deep transfer learning method to migrate the weighted model trained by the residual network on the Image Net dataset to the SAR image recognition work,which accelerating the time required for model training and improving the recognition efficiency.In addition,the addition of the Convolutional Attention Mechanism module further improves the recognition performance.Experiments are performed to verify the effectiveness of the algorithm by adding different degrees of random noise to the MSTAR data to simulate the environmental noise interference problem.Experimental results show that the recognition rate measured at 1% when noise level is 99.25%,which is 7.09% higher than that of the previous method.And the recognition rate measured at 15% when noise level is 94.63%,which is 38.85% higher than that of the previous method.These results show that the algorithm effectively reduces the target recognition influence of noise interference on SAR image.Finally,in terms of practical application,a SAR ATR algorithm based on YOLOv5 model migration in complex environments is proposed,because the detection and recognition effect of multi-target and small object detection in SAR images is poor in large scenes under the interference of complex environments(such as trees,buildings,etc.),the algorithm migrates the detection and recognition model trained by the YOLOv5 algorithm on the COCO dataset to the SAR image dataset,and uses the model to do the further training for achieving an effective detection and recognition effect.The experiment mainly integrates the large scene complex environment image with target image in the MSTAR dataset to obtain the SAR image object detection and recognition dataset for training.Experimental results show that the mean average precision(m AP)is 98.1% after migrating the YOOl Ov5 training model to the SAR image.This result has a high improvement over the previous algorithm result,so the SAR ATR algorithm has the practical effect. |
